LF-Net: Learning Local Features from Images

This repository is a tensorflow implementation for Y. Ono, E. Trulls, P. Fua, K.M. Yi, "LF-Net: Learning Local Features from Images". If you use this code in your research, please cite the paper.

comparison

Important Note regarding the use of ratio tests

Do NOT use the ratio test for descriptor matching! The commonly-used ratio test depends on the distribution of descriptor distances, and the threshold differs from one descriptor to another. Commonly used thresholds (0.9 0.7) are actually harmful for LF-Net. If you want to use the ratio test, you need to either tune this manually, or use statistical analysis as Lowe did for SIFT.

Installation

This code is based on Python3.6.5 and tensorflow with CUDA-8.0. For more details on the required libraries, see requirements.txt. You can also easily prepare this by doing

pip install -r requirements.txt

Docker image

We created a self-contained Docker image, for running the keypoint extraction demo easily. Make sure you have the nvidia docker runtime.

To launch a container:

docker run --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=all -ti --name lf-net-container -v /path/to/code_repo:/home jiangweivcg/lf-net-release-env

To run the run_lfnet.py script inside the container:

cd /home

python run_lfnet.py --in_dir=/path/to/images --out_dir=/path/to/outputs

Pretrained models and example dataset

Download the pretrained models and the scare_coeur sequence. Extract them to the current folder so that they fall under release/models/outdoor for example.

For other datasets, we do not plan to release them at the moment. Please do not contact us for explanations on the training phase. We are providing them as is as a reference implementation.

Updates new pretrained models

Download pretrained model without rotation augmentation

Updates since the arXiv version

The provided pre-trained models are trained with full 360 degree augmentation for orientation. Thus, the results you get from these models are slightly different from the one reported in arXiv. We have further included a consistency term on the orientation assignment.

Running the keypoint extraction demo

To run LF-Net for all images in a given directory, simply type:

python run_lfnet.py --in_dir=images --out_dir=outputs

In addition, you can easily do the 2-view matching demo through notebooks/demo.ipynb .

Training

Training code can be found in train_lfnet.py. We will not provide any support for the training process and datasets. All issues related to this topic will be closed without answers.